Deep Learning Model Restructuring for Variable Input Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning models face challenges in maintaining performance while being weight-lightened, as reducing the number of weights or biases often degrades their efficiency and computational load.
Innovation Solution
A method and system that change the structure of deep learning models by adjusting the size of feature maps and removing or modifying layers based on their output sizes, particularly those with 1×1 feature maps, to reduce computational load and memory footprint.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Weight of stationary object
If the number of weights or biases is reduced to lighten the deep learning model, then the model size is reduced, but the performance is degraded
Solution Approach 1:
The patent changes the resolution parameter of input data to dynamically adjust the structure of the deep learning model. By varying the input resolution, the model automatically adapts its internal feature map sizes and layer configurations, enabling weight-lightening at different resolutions while maintaining performance through structured design that preserves critical feature extraction capabilities across resolution changes
2Use of energy by moving object
If the resolution of input data is reduced to decrease computational load, then the computational load is reduced, but the model performance is degraded
Solution Approach 1:
The patent implements a dynamic model structure that adapts to different input resolutions. The model structure is not fixed but changes based on the resolution of input data, allowing it to optimize computational load for each resolution while maintaining performance through resolution-appropriate feature extraction. This dynamic adaptation enables the model to use fewer computations at lower resolutions without permanently degrading performance
3Device complexity
If layers with 1×1 feature maps are removed to simplify the model structure, then the model complexity is reduced, but the feature extraction capability is degraded
Solution Approach 1:
The patent applies local quality by selectively removing layers with 1×1 feature maps only in specific contexts where such layers do not provide critical functionality. The decision to remove layers is made based on analyzing the actual feature map sizes produced during inference, allowing the model to eliminate redundant complexity in certain local regions of the network while preserving essential feature extraction capabilities in other regions where 1×1 layers contribute meaningfully
Data Source
AI summary
Disclosed are a method and system for changing a structure of a deep learning model based on a change in resolution of input data. The method of changing a structure of a deep learning model may include generating, by the at least one processor, a plurality of input data having different resolution by performing various resolution changes on input data having given resolution, performing, by the at least one processor, inference on each of the plurality of generated input data through a deep learning model, checking, by the at least one processor, the size of a feature map output by each of layers included in the deep learning model while the inference is performed, and changing, by the at least one processor, the structure of at least one of the layers based on the checked size of the feature map.